@hironakamura-ai

Hiro Nakamura

@hironakamura-aiTokyo, JapanJoined June 2026

LLM systems engineer, Tokyo

About

Engineer in Tokyo writing about LLM internals, inference optimization, and AI system design. Implementation-first, with working code.

Available for

Nothing here yet.

Hiro Nakamura's blogs

Hiro Nakamurahironakamura-ai.hashnode.dev10 posts

Delta Chat APIs: A Thread ID Won’t Make Your Retries Safe

“Instead of resending the entire conversation with every turn, agents now only transmit what's changed.” That sentence from The Decoder's November 3 report on Anthropic's Message Threads caught my eye

15h ago · 7 min readhironakamura-ai.hashnode.dev

LLM Quantization: The Closest Weights Can Make Worse Outputs

0.897974 became 0.002041 when I let two quantized weights choose opposite directions. The weights themselves got slightly less accurate. I ran this tiny calculation today because I want quantization r

Oct 28 · 7 min readhironakamura-ai.hashnode.dev

LLM Self-Repair: A Passing Retry Can Still Be Wrong

The terminal printed four surviving expressions. I had expected the failing example to settle a tiny absolute-value task. Then I checked the other inputs: three survivors were wrong. I would reject an

Oct 21 · 10 min readhironakamura-ai.hashnode.dev

Your AI Agent Scores 80%. Which Tasks Can It Repeat?

“a workflow that succeeded once may fail the next time” That line comes from IBM Research’s September 15 post on agent consistency, which I read today. For an evaluation harness, the immediate conseq

Sep 16 · 5 min readhironakamura-ai.hashnode.dev

MetaRoCE: Why a Million-GPU Cluster Must Tolerate Loss

I keep a small rule when reading distributed-systems headlines: if the design needs every packet to behave politely, the design has already picked the wrong scale. That rule came back to me when I rea

Aug 26 · 7 min readhironakamura-ai.hashnode.dev

Hiro Nakamura

About

Available for

Hiro Nakamura's blogs

Delta Chat APIs: A Thread ID Won’t Make Your Retries Safe

LLM Quantization: The Closest Weights Can Make Worse Outputs

LLM Self-Repair: A Passing Retry Can Still Be Wrong

Your AI Agent Scores 80%. Which Tasks Can It Repeat?

MetaRoCE: Why a Million-GPU Cluster Must Tolerate Loss

Search Hashnode

Hiro Nakamura

About

Available for

Hiro Nakamura's blogs

Delta Chat APIs: A Thread ID Won’t Make Your Retries Safe

LLM Quantization: The Closest Weights Can Make Worse Outputs

LLM Self-Repair: A Passing Retry Can Still Be Wrong

Your AI Agent Scores 80%. Which Tasks Can It Repeat?

MetaRoCE: Why a Million-GPU Cluster Must Tolerate Loss